Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 148

square deviation from actual output to the ideal output. An unknown feature vector is
classified as belonging to a class, when deviation is less than a threshold. In general,
the threshold can be determined from the experimental studies.

7.5 Performance Evaluation with Different
Benchmark Database

In order to demonstrate the effectiveness of presented feature extractors and classi-
fiers, two sets of experiments are demonstrated in this section for relative performance
evaluation. In these experiments, the general performance of classifiers based on dif-
ferent neuron architectures are presented. The effect of using different statistical
feature extraction techniques in real and complex domain is also demonstrated with
different classifiers. In all experiments, four images of each subject are taken for
training and remaining images are considered for testing. The performance analysis
results presented in this chapter are average outcome of five training processes with
different weight seedings.
In order to evaluate the performance of different techniques, three standard bio-
metric measures are used. Recognition error rate (in other way percentage accuracy)
is the number of misclassification divided by total number of test images. False
acceptance rate (FAR) is the number of unauthorized persons considered as autho-
rized divided by total number of unauthorized attempts. False rejection rate (FRR)
is the number of authorized persons considered as unauthorized divided by total
number of authorized attempts. FAR and FRR are inversely proportional measure-
ments, therefore variable threshold setting are provided for users to keep balance. It
is generally desirable that recognition system must reject unauthorized attempts as
much as possible and perform well for authorized attempts. Thus, system must be
much stricter to unauthorized persons and slightly stricter for authorized persons. In
all experiments, our selection of threshold is inclined in a direction where we can
obtain significantly lower FAR and slightly higher FRR.

7.5.1 Performance in ORL Face Database

The first database considered for evaluating performance of proposed technique is
AT & T laboratories face database (formerly ORL) of Cambridge University [ 62 ]. It
contains facial images of 40 different subjects (persons) and 10 images per subjects.
None of the ten samples of each subjects is identical to each other. The images
are taken against a dark homogeneous background. There are variations in facial
expressions (open or closed eyes, smiling or frowning face), facial details (with or
without glasses, hairs), scale (up to 10%), orientation (upto 20 ◦ ). Each image was
digitized and presented by 92 × 112 pixel array whose gray level ranged between